Index · Blog

Notes on the work.

Mostly research notes. Pipeline choices, why they matter, what broke first time.

1. 8

   2026-08-27

   4,832 models on two machines that do not talk to each other

   The XAI 2026 architecture sweep trained 486 configurations per participant across a SLURM GPU cluster and an IBM Power9 CPU server. The models were the easy part. The logistics were the project.

   Code

   eeg · microstates · hpc · slurm · reproducibility

2. 7

   2026-08-27

   Never let the LLM set the price

   Two production LLM systems, one rule: the model narrates, the code decides. Pricing, risk flags, and anything else that must be right belongs to deterministic logic.

   llm · production · guardrails · gijima · curnance

3. 6

   2026-08-27

   The payout that raced the ledger

   A live withdrawal-fraud incident at a fintech taught me more about money movement than any design doc. The fix was not more validation. It was changing the order of operations.

   fintech · curnance · postmortem · mysql · idempotency

4. 5

   2026-06-26

   Do brain states actually repeat?

   A story from my PhD: getting a machine to rediscover the brain's "atoms of thought" on its own, and what it told me about whether those states genuinely recur.

   eeg · microstates · machine-learning · neuroscience · phd

5. 4

   2026-06-20

   Learning EEG microstates with a variational autoencoder: representation, the explained-variance gap, and the locus of recurrence

   A methods note on a Conv-VaDE pipeline for EEG microstates: why topographic images, why the model trails modified k-means on explained variance, the invariance of sequence statistics across deep variants, and the distinction between recurring patterns and recurrent dynamics.

   Code

   eeg · microstates · vae · vade · recurrence · rqa · modkmeans

6. 3

   2026-05-15

   EEG preprocessing, stage by stage

   Walking through the seven preprocessing stages that take a raw LEMON recording from disk to the input my Conv-VaDE actually sees.

   Code

   eeg · preprocessing · mne · pycrostates · lemon

7. 2

   2026-05-15

   Four ways to score the same model

   Cluster-validity scores disagree on the same model. So do the spaces you evaluate them in. The four-quadrant framework keeps both axes honest.

   Code

   eeg · microstates · clustering · evaluation · silhouette · davies-bouldin

8. 1

   2026-05-15

   The four corners of every topomap are not brain signal

   Every 40x40 topomap in the pipeline has a constant-zero ring of padded pixels outside the inscribed head-disc. Every loss and metric in the project masks them out. Here is why.

   Code

   eeg · microstates · topomap · preprocessing · metrics